Repository navigation
Release 1.1.0
Highlights
- LLM gateway — a single egress point for every LLM-backed engine. Translates
each provider's dialect (model naming, token-limit field, Anthropic's top-level
system, paired-field integrity) and classifies every outcome with a stable
code, so a refused request shape is distinguishable from a slow one. - CPU-aware sandbox scheduling — concurrency is chosen from queue depth and
bounded by the container's cgroup CPU quota, not by host cores. - Threat intel — signature-verified offline IOC import enables the intel
matcher tier (domain / IP / hash). - Pagination — reviews, inventory and reeval listings are windowed at 50 rows
and return a total. - Engine console — split into control (enable/disable) and status (last
result, duration, window counts).
Correctness
- Cache identity now covers the sandbox tier: a refreshed vulnerability database,
an edited rule, a bumped vendor pin or a different LLM model all re-judge
previously scanned packages instead of serving a stale verdict. - Verdict monotonicity — adding a confident finding can no longer relax a verdict.
- A Vault failure between scoring and signing no longer strands a scan.
- Least-privilege database grants are verified by authenticating as each module
account, not by asking whether the account exists.
Notes
Malicious-package advisories from OSV are now classified CRITICAL and titled
distinctly from CVEs. The aig adapter no longer records "no risk found"
placeholders as findings.
Full changelog: v1.0...v1.1.0